Evolutionary computation is frequently used in intelligent optimization and machine learning, but this type of machine learning differs significantly from the deep learning commonly discussed. Evolutionary computation has considerable potential in the brain-body integration design of robots. It may well be the next hot topic in artificial intelligence. This article is compiled from the speech given by Professor Yao Xin, IEEE Fellow and Head of the Department of Computer Science and Engineering at Southern University of Science and Technology, at the 2018 Shenzhen International Forum of Academicians on Robotics and Intelligent Systems, entitled "Why Study Evolutionary Computation?"
Many people may not be familiar with the Department of Computer Science and Engineering at Southern University of Science and Technology. The department was established in August 2016, and in 2017 we had our first batch of officially recognized undergraduate graduates. Last year, we also enrolled 19 master's students and 21 doctoral students. A year and a half later, we now have 19 teachers, and we plan to reach 55 in the future.
The research areas of the Department of Computer Science and Engineering at Southern University of Science and Technology (SUSTech) are divided into five main areas: Artificial Intelligence, Data Science, Theory, Systems and Networks, and Cognitive and Autonomous Systems. The Artificial Intelligence group has five professors, myself included. The other professors come from different places and have diverse backgrounds, but all are related to computational intelligence and neuroevolution. In addition, we have several PhDs from various parts of the country. We mainly engage in research on machine learning, optimization, and their intersections. Learning without optimization is incomplete; the purpose of learning is to make decisions, so learning and optimization need to be combined. Optimization considers many aspects, such as multi-objective optimization, dynamic optimization, and optimization in uncertain environments. Machine learning focuses on machine learning, online learning from data streams, and imbalanced learning. Another research group of ours focuses on Cognitive and Autonomous Systems, which involves both hardware and software. The hardware includes drones and swarm robots; the software is software robots.
Why study evolutionary computation?
First, let's look at what evolutionary computing is. I don't know how many people still write their own programs these days. Writing programs is like eating stinky tofu—you either love it or hate it. If you love or hate writing programs, how does it feel? Even though computers and robots are so intelligent now, you still type away furiously. Why? Because when you write a program, even if you're missing a comma or parenthesis, the editor always makes mistakes. Everyone who's written a program knows that spaces sometimes compile differently, causing compilation errors, which is very frustrating. Artificial intelligence has advanced to the point where humans rarely make the same mistake twice; such people are beyond help. But computers are sometimes very troublesome. You might say it's just missing a comma, but computers don't actually understand that. Therefore, computers are very inflexible and very fragile.
Furthermore, the standards and capabilities are relatively inferior. How many people are still using computers from ten years ago? No, computers are replaced every three years. As people age, they become increasingly intelligent, and robots and computers will inevitably need to be replaced every few years. The development path of machines is not quite the same as that of humans. Natural systems can be viewed as computational systems. In comparison, natural systems possess some truly remarkable characteristics that current computer science cannot yet replicate, such as self-healing capabilities and adaptive abilities. Nature offers many valuable lessons for computer scientists, so drawing inspiration from nature is not exclusive to computer science. The engineering community also frequently draws inspiration from nature. For example, in aircraft, the evolution from bird flight to biplanes and propeller planes all draw inspiration from bird flight.
Why draw inspiration from nature? Problem-solving methods for natural systems complement those for computers, and the solutions are often relatively simple, not overly complex. This is why evolutionary computation is studied. In reality, this research is far from simple, extending beyond evolutionary computation to include deep learning neural networks. Artificial neural networks are also inspired and influenced by the brain; evolutionary computation, or evolutionary algorithms, are derived from biological evolution. All living organisms in the world have evolved, following certain underlying patterns. Even finding just one ten-thousandth or one thousandth of these patterns can potentially lead to different approaches in computer design. This is one of the starting points for evolutionary computation.
What does evolutionary computation do?
The following four examples illustrate this.
The first example, which those interested in machine learning call machine learning, is actually data-driven model building. You're given a large amount of experimental data, and you're asked to identify the abstract model behind that data—you find the patterns in the data. Take aluminum alloy design as an example. Computer science is a fascinating field. In the past, aluminum alloy design involved modeling; now, we use evolutionary algorithms to minimize time spent in the lab. Let's say we know how to synthesize this aluminum alloy. We can conduct preliminary experiments in the lab, applying tension and pressure, and observing the deformation of the ingot after the pressure is released. The experts conducting the experiments can write a set of equations—here, four equations. This example contains six material constants. These material constants are values that everyone designing new materials particularly wants to know, but we don't actually know them. For material scientists, these values represent the material itself; for computer scientists doing mathematics, they are like variables. Now, I give you the equations, the experimental data, and the behavior of the material in the experiment. Can you find the material constants? This is somewhat like solving equations. What does solving these equations have to do with computer science? What does it have to do with evolutionary algorithms? This equation has no analytical solution; it can only be solved numerically. One approach is to transform it into an optimization method. When searching for a numerical solution, we look at the difference between the left and right sides of the equation. If the difference between the two sides is zero, we've found the solution. How do we find it? I buy a software package—one of the best in the world, from Oxford University—initialize the initial values of the six variables we're looking for, and then input the evolutionary algorithm. You'll find that the result you get for this problem is always the same as the initial values you input. This is because many numerical software packages make assumptions that don't hold true in real-world problems. Under these conditions, you can use evolutionary algorithms for optimization. This optimization involves solving the equation and finding the numerical graph. The solution found here is the best for designing aluminum ingots and aluminum alloys—it provides the most accurate numerical constants and material constants.
The second small example relates to optimization, which often involves an unwritten assumption: the optimization environment and the target are static. However, in reality, these conditions change. Imagine being given a goal and metrics, only to have the metrics suddenly changed halfway through, declaring it unacceptable. This is a real possibility. For instance, in northern regions during winter, roads must be salted to thaw the ice. Different countries have different laws; for example, in the UK, a law mandates salting road A if the weather forecast predicts a surface temperature below 2 degrees Celsius within two hours. Now, consider a specific problem: I have a convoy and the road network of road A. How do I dispatch the trucks to traverse the entire network within two hours? The various constraints involved make this a problem that cannot be found in operations research or mathematics textbooks. Mathematically, we assume the truck capacity is Q. However, the problem we face involves a government, a small convoy of only 11 trucks, and road conditions that cannot be assumed to result in uniform speeds, especially in winter. You have 11 cars and you need to schedule them. Half of them will leave you with 10 cars, and one car will break down. But you still need to complete the task. How can you complete it dynamically? Mathematical optimization methods can no longer solve this problem. Instead, we can use evolutionary algorithms, which are inspired by nature, to solve this complexity problem.
The third small example is about multi-objective optimization. "More, faster, better, cheaper" was a slogan from the 1970s that everyone liked to say, but it was difficult to achieve. To achieve more, faster, better, and cheaper, you can't just look at one indicator; you have to satisfy several indicators at the same time. As a decision-maker, you need to choose various compromise solutions. This is a very typical scenario of multi-objective optimization.
A concrete example is autonomous driving systems. Autonomous driving systems truly rely on software control. Whether you buy a software system or develop one yourself, you must prove that the system is correct. This requires software validation. If validation fails, software testing is necessary, trying various environments. However, a key point is that it's impossible to test every possibility. With limited resources and time, testing each module of a large software system to maximize its accuracy, and allocating limited personnel and funds to the major modules with the goal of achieving the highest overall software accuracy, all require evolutionary computation methods.
What is the relationship between evolutionary computing and robotics?
Finally, let's discuss the relationship between evolutionary computing and robotics. Within this field, there's a branch called evolutionary robotics. All living things in the world, including the human brain, have evolved. If the biological brain can evolve, why can't robots?
One advantage of using an evolutionary approach to robot design is that it allows for the simultaneous design of both the robot's control system and its morphology. Most of the time, research in these two areas is conducted by people from different disciplines—some focusing on machine learning, others on robotics—but in reality, the groups designing the controller and the control morphology should be combined, not separated.
Let me demonstrate a simulation experiment. This is an artificially constructed swimming line graph. The graph is segmented, each segment identical except for a head on the left and a tail on the right. My goal is to design or evolve a line graph that can swim from point A to point B, as quickly as possible. However, I won't tell the line graph how to swim. I don't provide any additional information on how it should swim; I'm simply giving you the task: swim from A to B in a straight line, as fast as possible. How is this line graph controlled? Each segment controls the movement of neurons, which is related to the position of neurons in the neural network. The line graph has circular holes, and the neurons are distributed at the physical locations of these holes. I've evolved a line graph that initially knew nothing, simply moving from A to B, and then evolved a controller on the neural network. In another experiment, I artificially shifted the left side of the line graph to the right—not intentionally, but manually—while the right side shifted to the left, becoming crooked. In this experiment, the same task was swimming from point A to point B, with the same goal: swimming as fast as possible. The method used was evolution. It involved slow swimming, with the fastest swimmers surviving and passing on their skills to the next generation. I wanted to see how the neural network controlled the two sets of experiments. You'll find it very interesting. In the first case, with the line graph unchanged and the initial straight shape, the evolved neural network is shown in the lower right corner, where the red dots represent the nerve positions—very symmetrical. Humans didn't tell it anything; the only feedback was swimming fast or slow. The upper left corner shows the initial line graph, where swimming was random, initially just swirling in the water. After a long time, around generation 1190, it found a very fast swimming pattern. In the second case, I changed the line graph shape, using the dotted line to represent it. The structure of this neural network perfectly compensated for this morphological deficiency, allowing it to see whether the swimmer veered left or right.
There are three very simple principles here: First, different forms require different neural networks. If you're designing a robot controller, you must consider whether it uses wheels or legs, two wheels, three wheels, four wheels, or eight wheels; if it uses legs, two legs or eight legs, and whether the legs are on the head or the bottom. This form is closely related to how you design your neural network. Second, this evolutionary process is designed as a whole, and a system designed as a whole is more effective than one designed separately. Third, the entire process is driven by evolutionary algorithms; it doesn't involve data, a million images or videos, making it a very different approach.
In summary, three points need to be made: First, evolutionary computation is frequently used in intelligent optimization and machine learning, but this type of machine learning is quite different from the deep learning commonly referred to. Second, evolutionary computation should have considerable applications in the brain-body integration design of robots. Third, evolutionary computation may be the next hot topic in artificial intelligence. People generally want to do as little work as possible but achieve the highest returns, and evolutionary computation is precisely a technology that can achieve this goal. It doesn't require data or labels, nor does it require correct or incorrect instructions; simply assign the task, and it can perform evolutionary computation.
